Corpus visualisation


This dashboard displays a research on a corpus consisting of songs added by students. The songs can be created by themselves, created with the help of AI or just plucked from the internet. My songs are AI generated with Jen. The first song is inspired by the movie ‘The Brutalist’. This movie involves a lot of ambient dystopian sounds and music. I really liked the sounds but I wanted more energy in them. So i prompted “dystopian breakbeat for brutalist buildings”. The second song was just whimsy thought inspired by Musks and Trumps plan to go to Mars. My prompt for it was “marching music for on mars”.

The class corpus is displayed in a scatterplot. This plot shows the emotional characteristics of all tracks. Each track is placed according to how positive/negative (valence) and how intense (arousal) it feels. Triangles represent tracks made by AI and circles represent tracks made by students or found online. AI tracks mostly fall in the middle, showing moderate emotions. The colors represent how instrumental (non vocal) a track is. Tracks with vocals often feel more energetic. My two AI-generated tracks get categorized as half instrumental but in reality they are fully instrumental. They do fall within the emotional range. The tracks don’t posses a lot of emotion. This remarks a characteristic of AI-generated tracks. People often talk about AI-generated tracks not having a soul. This dashboard will conduct further research on the aspects of AI generated tracks like rhythm and tonal structure.

Chroma features

Chromagram


This chromagram visualizes how pitch evolves over time. Notice the horizontal band around the pitch D. This indicates the tonal stability of AI-generated song. Another characteristic is the break around 50 to 75 seconds. In this period the pitch almost disappears. This matches in the song with the dropping of ambient sounds and only remaining percussion.

Cepstrogram


The cepstrogram visualizes the timbral characteristics of the audio trough Mel-frequency cepstral coefficients. The gradient indicates the variation of timbre. Here it is visible that AI-generated music has consistent but gradual shifts.

Chroma-based Self_similarity Matrix


The chroma-based SSM provides insights into repetitive structures in pitch. Here it is notacible that the song has clearly defined repetitive structures. This illustrates AI’s methodical pattern generation. This contrasts with more organic development found in human compositions.

Keygram


For my second song called “Marching Music for Mars” I chose a keygram instead of a chordgram. A keygram shows the overall tonal center of a piece over time. This works well for ambient music that does not have over clear chord changes.

I computed the keygram using the Euclidean norm and cosine distance. This combination made the pitch information more clear. The x-axis shows time and the y-axis shows different key templates. The color indicates how well the songs pitch content matches each key.

One clear feature is the black vertical strip between 50 and 75 seconds. This stripe appears because during that period the ambient sounds stop and only the percussion remains. Percussion has no clear pitch and so the keygram cannot match it to any key. For the rest the keygram shows that the song has a consistent tonal character in the ambient sections.

Novelties

Energy novelty function


This tab displays three different visualizations related to the tempo of the second song. The first plot is the energy novelty function. An energy novelty function has high peaks when the songs content changes abruptly. This energy novelty function reveals an abrupt change around 50-75 seconds. This resembles the same area like already indicated in the keygram. This indicates the dropping of ambient sounds with only the percussion remaining.

Spectral novelty function


The timbre based novelty function does not show any peaks. This suggests that the timbre qualities of the AI generated song change gradually.

Non-cyclic tempogram


The non-cyclic tempogram uncovers the underlying rhythmic structure by the clear horizontal lines. These indicate a steady pulse. This reflects the AI’s mechanical approach of generating music. But since this song was created with the prompt “marching music for on mars” this rhythmic aspect of the song could also reflect the AI effectively captured the concept of marching music.

Heatmap


This tab shows a heatmap. A heatmap displays a clustering of the class corpus based on five musical features: arousal, danceability, instrumentalness, tempo, and valence. Each row corresponds to a track in the corpus and each column represents one of the features. The dendrograms on the axes illustrate how songs and features according to their similarity get clustered.

I have picked these features for the following reason: - arousal and valence: tehy help describe how energetic and positive/negative a track sounds - danceability: useful for rhytmic qualities in music - instrumentalness: helps distinguish between voacl and non-vocal songs - tempo: captures the speed of music

The plot displays the following interesting aspects: - The upper-right group of tracks appear to cluster by having similar tempo. - The lower-left group of tracks cluster because they share the same valence and arousal levels. - Tempo and valence/arousal do not to seem to be correlated in the class corpus. This suggests that a faster tempo does not necessarily mean a more positive or energetic track.

Earlier analyses showed a gradual timbre change, a steady tempo and a mechanical feel. It is plausible that AI-generated songs might cluster in a region with medium to high danceability and instrumentalness (due to fewer vocal sections) and a stable tempo.

Conclusion

AI tracks generally show moderate emotional expression. This matches common belief. Most analyses in this dashboard focus primarily on one song, generated by AI. This song is “Marching Music for Mars”. This prompt could have already lead to a more monotone style. Nevertheless… This track shows stable pitch, gradual changes in sound quality and steady machinelike rythms. These traits highlight the skill of AI in creating music. The heatmap further shows how AI-generated tracks form their own group.